Abstract
Processing of local correlations is crucial to early vision, as it underlies identification of lines, edges, and texture. In natural scenes, correlations of low and high order are intertwined in a complex fashion. Here we extend an order-by-order analysis of sensitivity to image statistics to these complex combinations. To reduce dimensionality, we consider only configurations of binary pixels in 2x2 neighborhoods, as this captures informative local image statistics (Tkacik et al., 2010). The 16=2[sup](2x2)[/sup] configurations of binary pixels within this neighborhood correspond to 10 independent image statistics, of orders 1 to 4. In four subjects, we used a segmentation task to characterize sensitivity to these 10 image statistics, alone and in pairwise combinations. A perceptual metric with ellipsoidal iso-discrimination contours provides an accurate synopsis (~5% RMSE) of each subject’s sensitivities. Further, psychophysical data respect the symmetries of the checkerboard lattice: sensitivity to an image statistic is unchanged (within ~10%) following lattice rotations or reflections. These observations enable a decomposition of the 10-d space according to spatial symmetries. There is a 5-dimensional subspace comprising statistics of all orders, each of which is invariant under spatial rotations and reflections. This contains 5 of the ellipsoid’s principal axes; across subjects, axis directions agree within ~5%, and relative axis lengths agree within ~20%. There is a 2-dimensional subspace of statistics, each of which is sign-inverted by cardinal reflections. This contains 2 principal axes, with similar consistency across subjects. One of these axes is predicted to be metameric to random; this is supported by preliminary data. The last 3 principal directions correspond to other symmetries, and necessarily agree exactly across subjects. In sum, perception of local image statistics is highly conserved across subjects, and a Euclidean model built on an order-by-order analysis predicts, at least qualitatively, perception of complex combinations of image statistics.
Meeting abstract presented at VSS 2013